A judgment is delivered to the parties. If a party cannot read it, the delivery was formal rather than actual, and in jurisdictions where a substantial share of litigants do not read the language the court operates in, that gap is systemic rather than occasional.
AI translation of legal judgments in courts is how several judiciaries have chosen to close it, at a scale that would be impossible with human translation alone. India's Supreme Court built SUVAS, which translates between English and 19 Indian languages. By 2024 it had produced 36,271 Supreme Court judgments in Hindi and 17,142 across 16 other regional languages. Singapore's courts use translation trained on legal terminology.
Documents are only half of it. A court that translates its judgments and not its hearings has served the litigant who wants to read the decision and not the one who wants to know what was said about them, and the second is often the same person. Translation in a court covers written material and the spoken record, and the two behave differently enough that treating them as one workflow is where courts get into trouble.
The guide to writing a court AI policy treats translation of decisions as one of the workflows courts govern first. This article covers what makes legal translation different, for both.
Why courts translate, and for whom
The purposes are distinct and they imply different quality requirements, which courts frequently fail to separate.
Translation for a party, so they can understand a decision affecting them. This is the access-to-justice case and the volume driver.
Translation for the public, where decisions are published in more than one official language as a matter of law or policy.
Translation for another jurisdiction, where a decision is relied on in cross-border proceedings.
And translation for internal use, where a court needs to understand foreign material.
A translation adequate for a party to understand what happened may be inadequate as a version relied on in later proceedings. Deciding which purpose a translation serves determines what verification it needs, and treating them all identically means either over-investing in the low-stakes cases or under-investing in the high-stakes ones.
Cutting across all four is what is actually being translated. A filed document and a recorded hearing are different source material with different failure modes, and a court's translation policy needs to say which it covers.
Translating a recording is not translating a document
A document arrives as text. A proceeding arrives as audio, and translating it means transcribing it first, which stacks two error surfaces rather than one.
That stacking is the property to design around. A word misheard at transcription becomes a confidently mistranslated word downstream, and nothing in the output signals that the error entered two steps back. Reviewing a translated transcript against the source-language transcript catches translation errors; only going back to the audio catches transcription errors. Courts that review one layer and not the other are checking the cheaper half.
Interpreted testimony is the case that most needs a decision made in advance. Where a witness answers in one language and a court interpreter renders it in another, the recording already contains both, and machine-translating that produces a translation of a translation. Where the accuracy of the interpretation is itself in issue, which is not rare, that distinction is the entire question. The attribution scheme has to keep the witness and the interpreter apart rather than treating the exchange as one voice.
Two properties have to survive the crossing, and frequently do not. Speaker attribution, because a translated transcript that has lost who said what is a document about a hearing rather than a record of one. And timing, because a translated transcript stripped of its alignment can no longer be cited to a moment in the recording, which is what a reviewing court will want. Ask for a translated transcript in the export format during evaluation and check both.
The honest characterisation of the artefact follows from all this: a translated transcript sits two derivations away from the proceeding. It is a reasonable aid to understanding and a poor candidate for the authoritative version, and it should be labelled as what it is.
Where general translation fails on legal text
General-purpose translation performs well on ordinary prose and poorly on legal language, for reasons that are structural rather than a matter of quality.
Terms of art carry precise meanings within a legal system and frequently have no equivalent in another. A superficially corresponding term may denote something materially different, and a literal rendering produces text that reads authoritatively while conveying the wrong thing. That is worse than an obviously awkward translation, because nothing signals the error.
Legal drafting also uses constructions that resist translation: long sentences with nested conditions, deliberate repetition where ordinary writing would use a pronoun, and formulaic phrases whose meaning is conventional rather than literal. Systems tuned for fluency will smooth these, and smoothing legal text can change its effect.
Citation and reference formats differ by jurisdiction and are frequently mangled. A translated judgment whose citations no longer resolve is less useful than an untranslated one to a reader who could have looked them up.
Terminology control is the main lever
The practical response is domain adaptation rather than better general translation.
That means a maintained glossary of terms with their approved renderings, applied consistently, and it means treating that glossary as court infrastructure rather than a project artefact. Singapore's approach, training on domain-specific terms, is the pattern.
Consistency across a corpus matters as much as accuracy in a document. The same term rendered three ways across a body of translated decisions makes the corpus unusable for anyone trying to follow a line of reasoning. Where a court is also anonymizing decisions, the designation register and the terminology glossary have to be applied in a defined order, since substitution interacts with translation. That process is covered in anonymizing judgments before publication.
Verification, and what the translated version is
The question courts must answer explicitly: what is the legal status of a translated judgment.
The common positions are that the original is authoritative and the translation is provided for convenience; that both language versions are equally authoritative, which is the position in some multilingual jurisdictions; or that the translation is authoritative for a defined purpose and not otherwise.
Each implies different verification. A convenience translation can carry a clear disclaimer and lighter review. An equally authoritative version requires professional certification, and machine output is a first draft toward it rather than a substitute for it.
Labeling is what makes this workable. A translated decision should state its status on its face, so a reader knows whether they are looking at the decision or at an aid to understanding it.
Where a court cannot staff verification in every language it publishes, a risk-based model, with certified review for consequential material and clear labeling of unverified machine output, is the honest approach. The broader obligation framework is covered in language access obligations for transcription and translation.
Publishing without doubling the archive
A court publishing in several languages multiplies its published corpus, and the operational consequences follow.
Every version needs a stable location and citation. Corrections have to propagate to all versions, and a corrected original with an uncorrected translation is a live problem. Search should ideally work across languages rather than requiring a reader to pick one. And retention applies to each version.
Courts that treat translations as attachments to the original rather than as separate documents avoid much of this, since the relationship stays explicit.
How VIDIZMO fits, on both paths
Because the two source types run through different systems, it is worth being specific about which does what.
For documents, VIDIZMO AI Intelligence Hub handles the publication workflow: optical character recognition over scanned and filed material, translation across the resulting text, and semantic search over the multilingual corpus, which is what lets a reader find a passage without first choosing a language. Source citation identifies where retrieved material came from.
For the spoken record, the path most courts want runs through VIDIZMO DEMS, where the proceeding and its transcript already live under custody. Transcription runs across 82 languages with speaker diarization, automatic translation covers the resulting transcripts across 50+ languages, and transcript export templates produce timestamped and translated layouts, which is what preserves the attribution and the alignment the section above says to check for.
Underneath both, translation and speech-to-text are platform capabilities running self-hosted by default, including in air-gapped deployments, so a court under a residency constraint can translate unpublished judgments and sealed proceedings without the material leaving its environment.
Where neither extends: no platform certifies a translation, which remains a human professional act, and none decides the legal status of a translated version, which is a court determination. Courts requiring certified translations should treat machine output as a first draft toward one.
Building the process
Decide the purpose and legal status of each translation category before automating anything. Build and maintain a terminology glossary, and treat it as infrastructure. Apply anonymization and translation in a defined order. Label every translated version with its status. Verify proportionately to consequence. And plan for correction propagation across versions, because that is the failure that damages trust fastest.
Request a demo to test legal-domain translation and cross-language search on a sample of your own decisions, and on a recorded hearing.
